CICBUAPnlp at SemEval-2016 Task 4-A: Discovering Twitter Polarity using Enhanced Embeddings

Proceedings of SemEval-2016, pages 145–148, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics CICBUAPnlp at SemEval-2016 Task 4-A: Discovering Twitter Polarity using Enhanced Embeddings Helena G´omez-Adorno, Grigori Sidorov Center for Computing Research Instituto Polit´ecnico Nacional Av. Juan de Dios B´atiz C.P. 07738, Mexico City, Mexico helena.adorno@gmail.com sidorov@cic.ipn.mx Darnes Vilari˜no, David Pinto Faculty of Computer Science Benem´erita Universidad Aut´onoma de Puebla Av. San Claudio y 14 sur C.P. 72570, Puebla, Mexico darnes@cs.buap.mx dpinto@cs.buap.mx Abstract This paper presents our approach for SemEval 2016 task 4: Sentiment Analysis in Twitter. We participated in Subtask A: Message Polar- ity Classification. The aim is to classify Twit- ter messages into positive, neutral, and nega- tive polarity. We used a lexical resource for pre-processing of social media data and train a neural network model for feature represen- tation. Our resource includes dictionaries of slang words, contractions, abbreviations, and emoticons commonly used in social media. For the classification process, we pass the fea- tures obtained in an unsupervised manner into an SVM classifier. 1 Introduction In this paper, we describe our approach for the Se- mEval 2016 task 4 “Sentiment Analysis in Twitter” subtask A (Nakov et al., 2016), where the goal is to classify a tweet message as either positive, neu- tral, or negative. The main goal of our approach is to improve the feature representation obtained by a well-known neural network method–Doc2vec (Le and Mikolov, 2014), using dictionaries of abbrevia- tions, contractions, slang words, and emoticons. Approaches based on neural networks for unsu- pervised feature representation (or embeddings) of- ten do not perform data cleaning (Le and Mikolov, 2014; Socher et al., 2011), considering that the net- work itself would solve the related problems. These approaches treat special characters such as ,.!?# and user mentions as a regular word (Le and Mikolov, 2014; Brigadir et al., 2014). Still, in some works which use embeddings a basic data cleaning process (i.e., stopwords removal, URL filtering, and removal of rare terms) improves the feature representation and, consequently, the performance of the classifi- cation task (Yan et al., 2014; Rangarajan Sridhar, 2015; Jiang et al., 2014). The problem with the content of social media messages is that they usually have a lot of non- standard language expressions (Pinto et al., 2012; Atkinson et al., 2013). Due to the short nature of the messages, most of the users use a large vocabulary of slang words, abbreviations, and emoticons (Das and Bandyopadhyay, 2011). Slang words are not considered as a part of the standard vocabulary of a language, and they are mostly used in informal messages, while abbreviations are shortened forms of a word or name that are used in order to replace the full forms. Emoticons usually convey the current feeling of the message writer. For this task we propose a preprocessing phase using the dictionaries that we previously built for the task of Authorship Atribution (Posadas-Dur´an et al., 2015). These dictionaries are useful for pre- processing and cleaning messages obtained from several social networks, such …

Content truncated to 500 words